def test(self, testloader, test_size, inception=None):
with torch.no_grad():
self.G.eval()
self.D.eval()
x_real, _ = iter(testloader).next()
x_real = x_real.to(self.device)
x_fake = []
if inception:
f_real = []
f_fake = []
for i in range(test_size // 100):
z = torch.randn(100, self.z_dim, device=self.device)
x_i = self.G(z) * 0.5 + 0.5
x_fake.append(x_i)
if inception:
f_real.append(inception(x_real[i * 100:(i + 1) * 100]))
f_fake.append(inception(x_i))
x_fake = torch.cat(x_fake, dim=0)
if inception:
f_real = torch.cat(f_real, dim=0)
f_fake = torch.cat(f_fake, dim=0)
fid = FID_score(f_real.cpu(), f_fake.cpu())
nn_real, nn_fake = nn_accuracy(f_real,
f_fake,
device=self.device)
else:
fid = -1
nn_real, nn_fake = nn_accuracy(x_real,
x_fake,
device=self.device)
return fid, nn_real, nn_fake
embedding_layer_name = 'l2norm'
elif args.loss == 'softmax_corr':
embedding_layer_name = 'softmax'
else:
embedding_layer_name = 'embedding'
if not args.no_progress:
model.summary()
batch_transform_kwargs = {
'embedding' : embedding,
'num_classes' : data_generator.num_classes if args.cls_weight > 0 else None
}
if args.loss.endswith('_corr'):
loss = utils.inv_correlation
metric = 'accuracy' if (args.loss == 'softmax_corr') or (args.embedding == 'onehot') else utils.nn_accuracy(embedding, dot_prod_sim = True)
else:
loss = utils.squared_distance
metric = utils.nn_accuracy(embedding, dot_prod_sim = False)
# Load pre-trained weights and train last layer for a few epochs
if args.finetune:
print('Loading pre-trained weights from {}'.format(args.finetune))
model.load_weights(args.finetune, by_name=True, skip_mismatch=True)
if args.finetune_init > 0:
print('Pre-training new layers')
for layer in model.layers:
layer.trainable = (layer.name in ('embedding', 'prob'))
embed_model.layers[-1].trainable = True
if args.cls_weight > 0:
par_model.compile(optimizer = keras.optimizers.SGD(lr=args.sgd_lr, momentum=0.9, nesterov=args.nesterov, clipnorm = args.clipgrad),
embedding_layer_name = 'embedding'
if not args.no_progress:
model.summary()
batch_transform_kwargs = {
'embedding': embedding,
'num_classes':
data_generator.num_classes if args.cls_weight > 0 else None
}
if args.loss.endswith('_corr'):
loss = utils.inv_correlation
metrics = [
'accuracy' if
(args.loss == 'softmax_corr') or (args.embedding == 'onehot') else
utils.nn_accuracy(embedding, dot_prod_sim=True)
]
if len(args.top_k_acc) > 0:
for k in args.top_k_acc:
metrics.append(
utils.top_k_acc(k) if (args.loss == 'softmax_corr') or (
args.embedding == 'onehot') else utils.
nn_accuracy(embedding, dot_prod_sim=True, k=k))
else:
loss = utils.squared_distance
metrics = [
'accuracy' if args.embedding == 'onehot' else utils.nn_accuracy(
embedding, dot_prod_sim=False)
]
if len(args.top_k_acc) > 0:
for k in args.top_k_acc:
if not args.no_progress:
model.summary()
callbacks = []
batch_transform_kwargs = {'embedding': embedding}
if args.init_weights and (args.init_epochs > 0):
print('Pre-training linear transformation')
for layer in model.layers[:-1]:
layer.trainable = False
model.compile(
optimizer=keras.optimizers.Adagrad(lr=args.init_lr),
loss=utils.devise_ranking_loss(embedding, args.margin),
metrics=[utils.nn_accuracy(embedding, dot_prod_sim=True)])
model.fit_generator(data_generator.train_sequence(
args.batch_size,
batch_transform=transform_inputs,
batch_transform_kwargs=batch_transform_kwargs),
validation_data=data_generator.test_sequence(
args.val_batch_size,
batch_transform=transform_inputs,
batch_transform_kwargs=batch_transform_kwargs),
epochs=args.init_epochs,
callbacks=callbacks,
verbose=not args.no_progress,
max_queue_size=100,
workers=8,
use_multiprocessing=True)
loss = utils.inv_correlation if args.loss == 'inv_corr' else utils.squared_distance
# Load pre-trained weights and train last layer for a few epochs
if args.finetune:
print('Loading pre-trained weights from {}'.format(args.finetune))
model.load_weights(args.finetune, by_name=True, skip_mismatch=True)
if args.finetune_init > 0:
print('Pre-training new layers')
for layer in model.layers:
layer.trainable = (layer.name in ('embedding', 'prob'))
embed_model.layers[-1].trainable = True
if args.cls_weight > 0:
par_model.compile(optimizer = keras.optimizers.SGD(lr=args.sgd_lr, momentum=0.9, nesterov=args.nesterov, clipnorm = args.clipgrad),
loss = { embedding_layer_name : loss, 'prob' : 'categorical_crossentropy' },
loss_weights = { embedding_layer_name : 1.0, 'prob' : args.cls_weight },
metrics = { embedding_layer_name : utils.nn_accuracy(embedding, dot_prod_sim = (args.loss == 'inv_corr')), 'prob' : 'accuracy' })
else:
par_model.compile(optimizer = keras.optimizers.SGD(lr=args.sgd_lr, momentum=0.9, nesterov=args.nesterov, clipnorm = args.clipgrad),
loss = loss,
metrics = [utils.nn_accuracy(embedding, dot_prod_sim = (args.loss == 'inv_corr'))])
par_model.fit_generator(
data_generator.train_sequence(args.batch_size, batch_transform = transform_inputs, batch_transform_kwargs = batch_transform_kwargs),
validation_data = data_generator.test_sequence(args.val_batch_size, batch_transform = transform_inputs, batch_transform_kwargs = batch_transform_kwargs),
epochs = args.finetune_init, verbose = not args.no_progress,
max_queue_size = args.queue_size, workers = args.read_workers, use_multiprocessing = True)
for layer in model.layers:
layer.trainable = True
print('Full model training')
# Train model
callbacks, num_epochs = utils.get_lr_schedule(args.lr_schedule, data_generator.num_train, args.batch_size, schedule_args = { arg_name : arg_val for arg_name, arg_val in vars(args).items() if arg_val is not None })
embedding_layer_name = 'l2norm'
elif args.loss == 'softmax_corr':
embedding_layer_name = 'softmax'
else:
embedding_layer_name = 'embedding'
if not args.no_progress:
model.summary()
batch_transform_kwargs = {
'embedding' : embedding,
'num_classes' : data_generator.num_classes if args.cls_weight > 0 else None
}
if args.loss.endswith('_corr'):
loss = utils.inv_correlation
metrics = ['accuracy' if (args.loss == 'softmax_corr') or (args.embedding == 'onehot') else utils.nn_accuracy(embedding, dot_prod_sim = True)]
if len(args.top_k_acc) > 0:
for k in args.top_k_acc:
metrics.append(utils.top_k_acc(k) if (args.loss == 'softmax_corr') or (args.embedding == 'onehot') else utils.nn_accuracy(embedding, dot_prod_sim = True, k = k))
else:
loss = utils.squared_distance
metrics = ['accuracy' if args.embedding == 'onehot' else utils.nn_accuracy(embedding, dot_prod_sim = False)]
if len(args.top_k_acc) > 0:
for k in args.top_k_acc:
metrics.append(utils.top_k_acc(k) if args.embedding == 'onehot' else utils.nn_accuracy(embedding, dot_prod_sim = False, k = k))
cls_metrics = ['accuracy']
if len(args.top_k_acc) > 0:
for k in args.top_k_acc:
cls_metrics.append(utils.top_k_acc(k))